development and application of reservoir models and

Development and application of reservoir models and artificialneural networks for optimizing ventilation air requirements

in development mining of coal seams

C. Özgen Karacan ⁎

National Institute for Occupational Safety and Health (NIOSH) Pittsburgh Research Laboratory Pittsburgh, PA 15236, United States

Received 28 August 2006; received in revised form 18 December 2006; accepted 10 February 2007Available online 20 February 2007

Abstract

In longwall development mining of coal seams, planning, optimizing and providing adequate ventilation are very importantsteps to eliminate the accumulation of explosive methane–air mixtures in the working environment. Mine operators usually try tosupply maximum ventilation air based on the capacity of the system and the predicted need underground. This approach is neithereconomical nor safer as ventilation capacity may decrease in time depending on various mining and coalbed parameters. Thus, it isimportant to develop better engineered approaches to optimize mine ventilation effectiveness and, therefore, to ensure a safer workenvironment.

This study presents an approach using coalbed methane reservoir modeling and an artificial neural network (ANN) design forprediction and optimization of methane inflows and ventilation air requirements to maintain methane concentrations belowstatutory limits. A coalbed reservoir model of a three-entry development section, which is typical of Pittsburgh Coalbed mines inthe Southwestern Pennsylvania section of Northern Appalachian Basin, was developed taking into account the presence andabsence of shielding boreholes around the entries against methane inflow. In the model, grids were dynamically controlled tosimulate the advance of mining for parametric simulations.

Development and application of artificial neural networks as an optimization tool for ventilation requirements are introduced.Model predictions are used to develop, train, and test artificial neural networks to optimize ventilation requirements. The sensitivityand applications of proposed networks for predicting simulator data are presented and discussed. Results show that reservoirsimulations and integrated ANN models can be practical and powerful tools for predicting methane emissions and optimization ofventilation air requirements.Published by Elsevier B.V.

Keywords: Underground mining; Coalbed reservoir modeling; Artificial neural networks; Mine ventilation; Reservoir simulation; Mine safety

1. Introduction

In development mining, continuous advance of mineworkings may cause variations in the amount and rate ofmethane emissions into the mine atmosphere, which

⁎ Tel.: +1 412 386 4008; fax: +1 412 386 6891.E-mail address: [email protected].

0166-5162/$ - see front matter. Published by Elsevier B.V.doi:10.1016/j.coal.2007.02.003

leads to ventilation requirements continuously changingto maintain gas levels below legal limits. If adequateventilation is not provided, this condition can result inelevated methane concentrations, which can increase therisk of an explosion or a frictional ignition that may leadto an explosion. In order to improve the safety ofunderground coal mines, it is important to have thepredictive capabilities to estimate methane emission

222 C.Ö. Karacan / International Journal of Coal Geology 72 (2007) 221–239

rates based on coalbed and mining parameters at aparticular stage in mining and to be able to optimize themine ventilation requirements.

In addition to improved worker safety, optimizationof ventilation air requirements can also provide a costbenefit. Historically, mines tended to supply a fixedvolume of air based on the projected maximum demand(Hardcastle et al., 1997). In today's economic environ-ment, ventilation can consume 30–40% of the electricityused in underground coal mining operations. It is impor-tant to optimize the ventilation requirements to reduce theventilation costs in underground coal mines.

As development mining progresses, entry length andthe number of cross cuts (leakage) increase, requiringadditional airflow to adequately ventilate the workareas. Furthermore, as development lengths increase, itmay become increasingly difficult to keep methanelevels under statutory limits by ventilation alone. One ofthe most effective approaches to drain excessivemethane from the coalbed before mining starts and toshield the entries against methane inflow (Brunner et al.,1997; Noack, 1998; Diamond, 1994) is to drillhorizontal boreholes in the coalbed. This approach hasbeen proven to be very effective. Horizontal drillingtechnique and its application to degasify coal seams aredocumented in the literature (Thakur and Davis, 1977;Thakur and Poundstone, 1980; Thakur, 1997). Howev-er, no predictive techniques and guidelines exist thatestablish the distance between the entries and horizontalwells for optimal shielding and to detail the timingbetween well drilling and development mining foroptimal degassing of the coal.

It is important to improve predictive and optimizationmethods to provide adequate ventilation air based on thecoalbed and mining parameters. Over the years,reservoir-modeling methods and simulators have beendeveloped that can realistically represent the complexphysics of reservoir flow mechanisms in coalbeds andgas production operations with diverse well completions(King and Ertekin, 1991). These simulators have beensuccessfully applied in various coal basins for coalbedgas recovery using both vertical and horizontal bore-holes (Ertekin et al., 1988; Zuber, 1998; Young et al.,1993). These models offer advanced predictive capabil-ities to simulate the development mining process and theprediction of methane inflow rates (Zuber, 1997) as wellas the subsequent determination of airflow requirementsbased on coalbed and mining parameters. However, asthe number of independent variables increases, modelsolution and analysis become increasingly difficult.

Artificial neural networks (ANNs), on the otherhand, are adaptable systems that can determine relation-

ships between different sets of data. ANNs have beendeveloped to solve problems where conventionalcomputer models are inefficient. These problems areeither non-polynomial types having no polynomialrelationship or very complex problems that are difficultto describe mathematically. The key advantages ofneural networks are their abilities to learn, to recognizepatterns between input and output space, to generalizesolutions, and to interpret incomplete and noisy inputs.Statistical techniques such as multiple regressionanalysis have been used widely for these kinds ofproblems, but they often fail in accuracy of prediction,especially in the face of highly non-linear relationshipsor incomplete and noisy data. Owing to the massivelyparallel, distributed processing nature of ANNs that canimprove their ability through dynamic learning, theymay outperform other methods of prediction andoptimization. Due to their inherent capabilities andflexibilities, artificial neural networks are a powerfulpredictive and optimization method when used inconjunction with reservoir simulation methods forpredicting methane inflow and optimizing ventilationrequirements during development mining.

2. Objective and description of the study

This paper presents the development and applicationof reservoir simulation and artificial neural networkmodels for improved prediction and optimization ofmethane inflow rates and ventilation requirementsduring development mining sections in coal seams.The reservoir models were developed based on a typicalthree-entry Pittsburgh Coalbed mine operating in theSouthwestern Pennsylvania section of the NorthernAppalachian Basin. These models were constructedwith and without the presence of degasification andshielding against methane. They were developed usingComputer Modeling Group's (CMG, 2003) composi-tional reservoir simulator (GEM). The models were run“dynamically” to simulate advance of entry develop-ment using a “restart” approach (Karacan et al., 2005, inpress). The reservoir model predictions were comparedwith the in-mine monitoring data of air quality and flowrate obtained during tailgate and headgate entrydevelopment around a longwall panel. Various miningand degasification-related parameters were consideredfor parametric runs using this reservoir simulator.

The “synthetic” cases generated while performingreservoir simulation runs were used to develop, train,and test the artificial neural network (ANN) models,which were developed using Neurosolutions 5.0 software(NeuroDimensions, 2006). The proposed ANN models

Table 1Values of some of the reservoir parameters of Pittsburgh Coalbed usedin the models

Parameter Value

Permeability-face cleat (md) 4Permeability-butt cleat (md) 1Effective porosity (%) 4Effective fracture (cleat) spacing (ft)/(m) 0.1/0.03Langmuir pressure (psi)/(MPa) 326/2.25Langmuir volume (scf/ton)/(cc/g) 490/15.5Desorption time (days) 20Initial water saturation (%) 60Coal density (lb/ft3)/(g/cc) 84.7/1.35Pressure (psi)/(MPa) 90/0.61

223C.Ö. Karacan / International Journal of Coal Geology 72 (2007) 221–239

were evaluated for their learning and predictive perfor-mances of reservoir simulation results.

3. Reservoir-modeling approach to simulatedevelopment mining

3.1. The mining site monitored for model comparisons

For modeling development mining, a three-entrytailgate and headgate development, typical of coal minesoperating in Pittsburgh Coalbed in the SouthwesternPennsylvania section of the Northern AppalachianBasin, was analyzed (Fig. 1).

During the mining of headgate and tailgate entries,ventilation air qualities and flow rates were measured bythe operating mining company using methanometersand flow meters at the monitoring locations shown inFig. 1. Methane inflow rates were quantified based onthe measured data. The data was reported as averagemonthly concentrations of methane, airflow rate, rawand clean tonnages of produced coal, and total lineardistances advanced during mining. The length of eachentry section was around 11,000 ft (3353 m) and took8–9 months to mine.

3.2. Coalbed reservoir and modeling parameters

The base model for formulating development miningwas a coalbed methane reservoir model. A single-layercoalbed reservoir model was created in Cartesiancoordinates to model fluid flow in the unmined sectionsof the coalbed and in the entries, as well as to simulatedevelopment mining operation. In the development ofthe model, the parameters and their average values inTable 1 were used for the Pittsburgh Coalbed. The datafor pre-mining reservoir properties were gathered fromprevious NIOSH publications, external reports, personalcommunications with the operating mining company,

Fig. 1. The modeled long wall panel showing entry sections where mini

and previous history matching studies (Karacan et al.,2005, in press).

In the Pittsburgh Coalbed in the Central and NorthernAppalachian Basin, face and butt cleats are perpendic-ular and parallel, respectively, to the fold axis. At thestudy mine, face cleats were oriented in the E–Wdirection and butt cleats in the N–S direction. Thisinformation indicates that the direction of mine advanceis parallel to the face cleats and perpendicular to buttcleats. When positioning the simulation grids andassigning the permeabilities, attention was given tocleat directions.

The gas content and adsorption related data for thePittsburgh Coalbed was obtained from the results ofmethane adsorption and direct method of gas contentdetermination tests on various coal samples (Diamondet al., 1986) and from an unpublished report of Langmuirparameters of a site-specific core. The spatial distribu-tions of fracture permeabilities of the Pittsburgh Coalbedhave not been previously established or reported. How-ever, based on a large scale modeling study involvingestimation of some coalbed parameters (Karacan et al.,2005, in press), the average permeabilities for thePittsburgh Coalbed were estimated to be 4 md in the

ng and ventilation data was gathered during development mining.

Table 2The parameters and their range of values changed in simulation runs

Parameter Range

Mining height (ft)/(m) 5–7/1.52–2.13Entry length (ft)/(m) 1000–12,000/305–

3658Mining rate (ft/day)/(m/day) 25–175/7.6–53.3Methane concentration in mine air (%) 0.5–1.5Distance of shielding wells to entries (ft)/(m) 19–87/5.8–26.5Degasification duration before mining (days) 0–180

Fig. 2. A three dimensional snapshot of pressure distribution duringdevelopment of the coalbed. Pillars, ventilation scheme and themodeled shielding wells are also shown.


face cleat direction and 1 md in the butt cleat direction. Inthe simulations, these values were taken to be uniformthroughout the layer.

The gas relative permeability curve of the coalbedwas estimated by matching the average gas productionrate from the simulated boreholes in the models to thereported average methane production rates of threehorizontal degasification boreholes in the mining area[8.42 scf/day/ft (0.782 m3/day/m) vs. 8.70 scf/day/ft(0.807 m3/day/m)]. A similar field-data-based adjust-ment could not be made to the water relativepermeability curve because no data was available.Thus, the curve was set to low values so the wellswould experience low amounts of water [0.0016 bbl/day/ft (0.0015 m3/day/m)].

3.3. Modeling of development mining process andperforming the simulations

Various factors control the ventilation requirementsduring development mining. Among these factors,coalbed parameters and mining parameters are probablythe most important. In this study, the Pittsburgh Coalbedwas selected as the primary coalbed of interest and itsaverage properties (Table 1) were used in the models.Thus, the variables originating from coalbed reservoirproperties were mostly eliminated from parametric runsand subsequent analyses. In that sense, these modelsmay be considered site-specific to Pittsburgh Coalbed.However, the effects of coalbed parameters can beinvestigated easily using the same approach to increasethe predictive capability for other coalbeds. The mainareas of emphasis were mining parameters, the methaneconcentration level to be maintained, and the presenceor absence of a degasification/shielding program toreduce the amount of methane emissions into thedeveloped entries. Based on this approach, two differenttypes of models, with and without horizontal degasifica-tion wellbores, were developed and simulation runswere performed. The parameters and their ranges ofvalues are shown in Table 2.

For modeling driveage of tailgate and headgateentries, a three-entry development model around alongwall panel was studied. Fig. 2 shows a snapshot ofthe model that represents mining advance, pillar layout,and the ventilation scheme. An oblique plane wasremoved from the picture for a better visualization of thedifferent elements of the model. The middle entry wasmodeled as the “track” or haulage entry, where intake airwas entering. In this entry, ventilation air injection andmine pressure assignments to the grids were performedusing an injector well. The entry to the left of “track”was designated as the “belt” entry. The third entry wasdesignated as the “return” entry carrying away majority(N90%) of the methane emissions and the methane-loaded ventilation air. The producer wells used in theseentries both assigned the mine pressures and monitoredthe amount of methane in the produced gas stream. Thiswas the amount of methane entering the entries asmining progressed and was used to calculate ventilationrequirements to dilute it to desired levels at any stage inmining.

During development mining, the entries and the crosscuts constituting the volume to be ventilated arecontinuously extended imposing a moving boundary-type problem in modeling. As the continuous mineradvances, new volumes are created that must beventilated while new surfaces are created that liberategas into that volume. In this study, the development of athree-entry continuous mining model, shown in Figs. 2and 3, was handled using “restart” model runs. Thesemodels were run sequentially, each characterizing anadvance in entry development with a specified devel-opment rate where coalbed properties were replacedwith the assigned properties of the entries in the modelsand ventilation-related features are built. The simulation


outputs from the previous model run were written in a“restart” file and then used by the next model as theinput run, while reservoir parameters were changed tocharacterize the development of entries. This approachwas used recently by Karacan et al. (2005, in press) tomodel the longwall mining process and the perfor-mances of gob gas ventholes. Zuber (1997) approached

Fig. 3. (A–B) A portion of the grid model showing the advance of mining at tw

the problem of roadway development by using areservoir simulator and modifying the grids at appro-priate times at the differential equation level withoutrestarting the simulation.

In building the restart models, all three entries weredeveloped simultaneously to a specific distance in apredefined amount of time, allowing calculation of the

o successive stages and the resultant ventilation airflow paths (arrows).


rate of mining advance. For different mining rates, thetimes in the recurrent data set were changed. Thus, therates reported in this study are not linear mining rates,but instead represent the rates of the mine sectionadvance. Based on this approach, a 150 ft (45.7 m)section advance corresponds to 570 ft (173.7 m) oflinear mining distance in the model, including entriesand cross cuts.

As the entries were developed at the designated ratesin the models by changing coalbed properties, the pillarsand cross cuts were also developed at the same time. Thepillars between the entries were 125 ft (38.1 m) in lengthand 75 ft (22.9 m) in width, and had the same propertiesas the coalbed. However, during the development ofentries, each of which was 20 ft (6.6 m) in width and thesame height as the coalbed, the permeability wasreplaced with a high permeability [109 md (10−6 m2)]in all three directions. Also, the coal-matrix pressures inthe mined grids were assigned to atmospheric pressuresto simulate mining process.

The development of cross cuts was modeled the sameway as the entries. However, during each simulation runonly the last set of cross cuts was left fully open forventilation airflow. During the development of crosscuts, stoppings or block walls (between track-belt andtrack-return) were automatically created between therestart runs to force ventilation flow through the lastcross cuts at each section advance. A “curtain” resis-tance in the last section of the “track” diverted some ofthe intake air towards the “belt” entry to ventilate bothbelt and face (Fig. 3A and B). A permeability of 100 md(10−13 m2) was assigned to the stoppings to representleakage, a common occurrence in underground mining.Fig. 3A and B shows the progress of mining betweentwo successive steps. These figures show the entries,cross cuts, and the pillars created during simulationsplus the path of the ventilation airflow.

In this study, two different modeling approacheswere undertaken to predict methane inflows and toestimate ventilation requirements during developmentmining: with and without degasification wellboresaround the entries to shield them from migratingmethane. The variables of these models were miningrate, mining height (coalbed thickness), methane per-centage in the ventilation air, length of developedentries, pre-mining degasification duration and theproximity of boreholes to the entries, as summarizedin Table 2. In the models with shielding boreholes,the borehole diameters were 3 in. (7.6 cm), with nowellbore skin, and they were operated with −0.2 psia(−1260 Pa) bottom-hole pressure. The lengths of theboreholes were equal to the lengths of the entries.

The wellbores were operated during mining regard-less of the duration of the pre-mining degasificationperiod.

4. Analysis of reservoir model predictions ofmethane inflow and airflow

4.1. Description of mining conditions and parametersin the monitored mine and comparisons with modelpredictions

During mining of tailgate and headgate entries shownin Fig. 1, various parameters of mining and ventilationwere measured by the operating mining company. Thelinear advance distances, lengths of the exposed ribs, aswell as raw and clean coal tonnages were reported asmonthly totals. The measured air quality, ventilationairflow rates, and calculated methane inflow rates werereported as monthly averages.

The mining distances were reported as linear dis-tances mined (including entries and cross cuts), whichwere around 4500–6500 ft (1372 m–1981 m) permonth, rather than the daily advance rate of all threeentries in the direction of mine advance as formulated inthe model. In order to establish a comparison betweenmeasured data and the reservoir model formulation,linear mining distances were converted to net advance ofentries. This approach revealed a ratio of 0.27 (1 sectionlength=3.7 linear length) between these two lengthscales. Thus, the linear mining distances were convertedto distances in the direction of mine advance. Theadvance rates, using the lengths in the direction of mineadvance, were calculated based on two eight-hourproduction shifts and one maintenance shift per dayfor 6 days/week. The majority of calculated sectionadvance rates were between 70–110 ft/day (21.3–33.5 m/day) during production shifts and workingdays. The average rate of mining until both entrysections were developed to full length was approxi-mately 80 ft/day (24.4 m/day).

The Pittsburgh Coalbed generally has a uniformthickness of about 7 ft (2.1 m), although it can varyfrom 6–8 ft (1.8–2.4 m) in the studied region. Theseregional variations may affect methane inflow rateinto the mine and the methane concentration in theventilation air. The coalbed thickness information wasnot included in the data sheet of the monitoring study.Thus, raw tonnages and the average dimensions of theentries and cross cuts were used to convert this in-formation into mining height. The results indicatedthat the average thickness of the coalbed varied from6.2–7.6 ft (1.9–2.3 m) in the direction of mining. The

Table 3Measured (at the monitoring point in return entry shown in Fig. 1) and calculated parameters for mining of tailgate entries

Mining durationfrom start (months)

Airflow rate (cfm)— measured

Methane concentration(%) — measured

Mining advance rate (ft/day) — calculated

Length of mined entries fromstart (ft) — calculated

Coal thickness (ft)— calculated

1 44,425 0.250 103 1758 6.92 55,533 0.200 100 3468 7.33 52,965 0.425 77 4793 7.64 60,239 0.500 84 6222 6.95 67,850 0.675 42 6939 7.46 64,548 0.750 83 8358 6.57 64,725 0.925 75 9639 6.28 71,280 0.825 98 11,311 6.4


average thicknesses of the coalbed along the tail-gate and headgate sections were 6.9 ft and 6.8 ft,respectively.

Changes in reported average methane concentrationduring monitoring were caused by changes in miningrates, mining height, and ventilation rates. The measuredmethane concentration in the mine atmosphere in-creased from 0.25% at the beginning of the headgateand tailgate development to 0.85% at the end. Theaverages over the entire mining period were 0.56% and0.57% for tailgate developments and headgate develop-ments, respectively. However, it should be noted that theventilation airflow increased from 44,000 cfm (ft3/min)to 70,000 cfm (20.76 m3/s and 33.04 m3/s) measured atmonitoring points shown in Fig. 1 during this sameperiod (Tables 3 and 4). Although the mine wassuccessful in maintaining methane levels below the1% statutory limit for the whole mining period,increasing methane concentrations suggest that theincrease in methane inflow rate offset the increase inventilation airflow. To maintain a constant, low methaneconcentration, the ventilation airflow should haveincreased as the length of the developed entriesincreased. Tables 3 and 4 show the measured andcalculated parameters discussed above for mining oftailgate and headgate entries in Fig. 1.

Table 4Measured (at the monitoring point in return entry shown in Fig. 1) and calcu

Mining durationfrom start (months)

Airflow rate(cfm)-measured

Methane concentration(%) — measured

Mining a(ft/day)

1 47,830 0.425 802 47,949 0.750 783 45,530 0.475 634 49,445 0.325 205 53,065 0.433 906 43,915 0.600 967 63,540 0.725 898 69,120 0.850 109

4.2. Comparison of reservoir model prediction with in-mine monitoring data

The predictive performance of the reservoir modelwas compared with in-mine measurements of methaneinflow, reported as monthly averages, during develop-ment mining of the tailgate and headgate entries. Sincethere was no report indicating that the area wasdegasified using horizontal wells, reservoir modelsdeveloped without the degasification option and theiroutputs of predicted methane inflow and ventilationairflow rates were compared with the measured data.The reservoir models utilized mining parameters thatwere close to the calculated averages of monitoredvariables, i.e., 7 ft (2.1 m) mining height, 70 ft/day and110 ft/day (21.3 m–30.5 m) advance rate, and 2000–12,000 ft (610–3658 m) development length. Airflowrates were calculated to produce a constant 0.5%methane concentration based on methane inflowpredictions.

Fig. 4 shows the simulator predictions and measuredmethane inflows into the headgate and tailgate entries.The solid markers and the trendlines show the simulatorpredictions for methane emissions at two differentmining rates, upper and lower limits of the calculatedaverage advance rate of the mining, as a function of

lated parameters for mining of headgate entries

dvance rate— calculated

Length of mined entries fromstart (ft) — calculated

Coal thickness (ft) —calculated

1360 6.72697 6.63781 6.34131 6.85676 7.17319 6.78848 6.510,701 7.6

Fig. 4. Comparison of methane inflow calculated in the mine based on measurements of airflow rate and air quality and the simulator predictions attwo different mining rates.


entry lengths. Since the simulations use constant miningand coalbed parameters, the methane inflow resultsshow a continuous increase with the length of minedentries, as opposed to the scattered increase in measureddata. However, the developed model closely predicts therange of measured data.

The predicted airflow rates needed to maintain aconstant 0.5% methane level in the ventilation air andthe measured airflow rates that resulted in an averagemethane concentration of 0.57% are shown in Fig. 5.This figure shows that, because of methane inflowincreases, simulator predictions of air requirementincrease with entry length to keep the methane levelsat 0.5%. Although the average predicted airflow rate is

Fig. 5. Comparison of airflow rate measured in the mine and the simulatormethane level of 0.5%.

close to the average of the measured values, themeasured values at the start and end of the mining arehigher and lower than the predicted, respectively. Themodel calculates the required amount of air based on thesimulated methane inflow. During mining, the ventila-tion air rate was kept in a narrow range regardless of thelength of the developed section or the mining rate. Infact, methane levels increased from 0.20% at thebeginning of development to 0.85% at the end ofdevelopment (Tables 3 and 4), indicating an abundanceof ventilation air at the beginning and a lack of airflow atthe end. This comparison suggests that adjusting airflowquantity based upon mining progress using a predictivemethod may be safer and more economical than

calculation based on two different mining rates to maintain a constant

Fig. 6. The change of methane inflow rates predicted by the reservoir simulation as a function of the mining rate for different entry lengths.


supplying a fixed volume of air based on projectedmaximum demand.

4.3. Parametric analysis of reservoir-modeling results

The reservoir model was used to generate parametricsimulations to cover a range of values for the mining-related parameters shown in Table 2. Some of thoseresults, which later were used to develop an artificialneural network (ANN) prediction and optimization tool,are presented in this section to show the relationshipsand effects of different parameters on methane inflowinto the entries during development mining.

The reservoir models predicted the methane inflowrate into the entries. Once the methane inflow rate isknown, the airflow requirements can be calculated tokeep methane concentrations at a required level. Fig. 6

Fig. 7. Calculated airflow rates required to maintain various methan

shows the predicted methane inflow rates at 2000, 8000,and 12,000 ft (610 m, 2438 m, and 3658 m) of roadwaydevelopment in the 7-ft-high Pittsburgh Coalbed withdifferent development rates. Results show that themethane inflow rate increases with the length of thedevelopment section because of the increase in surfacearea of the exposed coalbed. This increase is a strongfunction of the mining rate. However, the effect of themining rate on methane inflow is less pronounced forshorter development distances than longer distances.For 2000 ft of roadway development, increasing themining rate from 25 ft/day (7.6 m/day) to 175 ft/day(53 m/day) increases the methane inflow rate from109,100 scf/day (3091 m3/day) to 214,000 scf/day(6062 m3/day), which is approximately a two-foldincrease. With an 8000 ft entry development, afterincreasing the mining rate from 25 ft/day to 175 ft/day,

e concentrations for different mining rates and entry lengths.

Fig. 8. Methane inflow rate predictions for different development lengths and various pre-mining degasification durations. “No degasification” in thefigure legend corresponds to the case where there was no degasification boreholes. On the other hand, 0 months pre-mining degasification andmethane production from the wells starts at the same time as the mining. Data represents a development rate of 70 ft/day (21.4 m/day) and horizontalboreholes located 19 ft (5.8 m) from the entries.


the methane inflow increases from 244,300 scf/day(6921 m3/day) to 700,500 scf/day (19,850 m3/day).Similarly, increasing the mining rate from 25 ft/day to175 ft/day in developing 12,000 ft of entries increasesmethane inflow rate from 309,000 scf/day (8755 m3/day) to 991,150 scf/day (28,079 m3/day).

The methane inflow simulations performed fordevelopment mining can be used for evaluating anddesigning ventilation requirements. Fig. 7 shows thecalculated airflow requirements as a function of themining rate to keep methane concentrations at pre-determined concentrations during development of 2000

Fig. 9. Effect of degasification duration on the methane inflow rate for vcorresponds to the case where there were no degasification boreholes locatedurations using boreholes located 19 ft from the entries.

and 12,000 ft (610 m and 3658 m) entry sections. Forthis calculation, the methane inflow rates shown inFig. 6 were used. With shorter development distances,only minimal adjustments to ventilation airflow areneeded to keep methane levels below 1%. Such ad-justments appear to be independent of the mining rate.With longer development distances, more ventilationairflow is needed to keep methane levels constant andthe amount is more strongly affected by the mining rate.If the ventilation rate cannot control methane levels,decreasing the mining rate can be considered as adesired control.

arious development lengths and rates. “No degasification” in X-axisd 19 ft the entries. Different months are the pre-mining degasification

Fig. 10. Effect of proximity of shielding wells to the entries and the length of entries on methane inflow rate. “No degasification” in the figure legendcorresponds to the case where there were no any degasification boreholes. The other data are after 6 months of degasification. The simulated datashown is for 70 ft/day (21.4 m/day) development rate.


Using in-seam horizontal boreholes is an effectiveapproach to shield entries against methane inflow duringdevelopment mining. To simulate the shielding effectsof degasification wells on methane inflow and on theventilation air requirements during development min-ing, pre-mining degasification periods of 0 months (i.e.,no pre-mining degasification) and 3 and 6 months weremodeled. The boreholes were placed 19–87 ft (5.8–26.5 m) away from the entries for parametric analysis. Inall cases, the wellbores were operated during miningregardless of the duration of pre-mining degasification.

Fig. 8 shows a comparison of methane inflow rates asa function of development length for horizontal bore-holes located 19 ft (5.8 m) away from the intake and

Fig. 11. Effect of mining rate and shielding well proximity when mining 20006 months prior to the start of development mining.

return entries and operated for various pre-miningdurations. The data represents a 70-ft/day (21.4 m/day)development rate in a 7-ft (2.1-m) thick coalbed.Simulations show that the methane emissions are highestwhen shielding is not used against methane inflowbefore or during mining. Emission rates progressivelydecrease as a result of shielding and degasification. Forinstance, even if the boreholes are not operated beforemining and begin to operate when mining starts, theinflow rate decreases about 25% compared to thecompletely unshielded case. The methane inflow ratesafter longer pre-mining degasification times are less.

Fig. 9 shows the effect of mining rate and pre-miningdegasification duration when mining development

-ft and 12,000-ft (610-m and 3658-m) long entries. Wells operated for


entries with lengths of 2000 and 12,000 ft (610 m and3658 m). The presence of methane shielding and theduration of pre-mining degasification are more impor-tant when development distances are long and miningrates are high.

Fig. 10 shows the effects of the proximity of theshielding wells to the entries and the developmentlength on methane inflow rate. The simulated datashown is for 70-ft/day (21.4 m/day) development rate in7-ft (2.1-m) thick coalbed after 6 months of pre-miningdegasification, except for the no-degasification case,which does not have any shielding. This figure showsthat those wellbores closest to the entries are mosteffective in reducing methane inflow rates duringdevelopment mining, about 50% at 12,000 ft (3658 m)compared to no-shielding. Thus, positioning wells asclose to the entries as practically possible serves betterfor shielding purposes.

Fig. 11 shows the effect of mining rate and shieldingwell proximity for mining 2000-ft and 12000-ft (610-mand 3658-m) long entries after operating the wells for6 months prior to the start of development mining.Operating shielding wells that are located at close prox-imity to the entries for 6 months prior to mining mayreduce the methane inflow rate into the entries as much

Table 5The performance results of different ANN topologies tested to find thedegasification or shielding boreholes

Hiddenlayers

Neuronperhiddenlayer

Epoch Transferfunction

Momentum Training

Minimum M

1 1 4 500 Tanhaxon 0.3 0.004622 1 4 500 Tanhaxon 0.5 0.004993 1 4 500 Tanhaxon 0.7 0.003124 1 6 500 Tanhaxon 0.7 0.001935 1 6 1000 Tanhaxon 0.7 0.000646 1 4 1000 Tanhaxon 0.7 0.001847 1 4 2000 Tanhaxon 0.7 0.000598 1 6 2000 Tanhaxon 0.7 0.000299 1 6 1500 Tanhaxon 0.7 0.0006210 1 6 2000 Tanhaxon 0.8 0.0004611 2 6 2000 Tanhaxon 0.7 0.0003512 1 4 500 Sigmoid 0.3 0.0120113 1 4 500 Sigmoid 0.5 0.0123714 1 4 500 Sigmoid 0.7 0.0100315 1 6 500 Sigmoid 0.7 0.0057416 1 6 1000 Sigmoid 0.7 0.0040717 1 4 1000 Sigmoid 0.7 0.0055818 1 4 2000 Sigmoid 0.7 0.0020219 1 6 2000 Sigmoid 0.7 0.0017120 1 6 4000 Sigmoid 0.7 0.0015321 2 6 2000 Sigmoid 0.7 0.01249

Data rendered in bold are the most suitable topologies in each category, amo

as 50% during development mining, especially whenmining longer entry sections at higher mining rates.

5. Development and application of artificial neuralnetworks (ANNs) integrated with reservoir simulationas a prediction and optimization tool for developmentmining

5.1. Basic principles and components of an ANN

A neural network is a computing tool that processesinformation by its dynamic state response to externalinputs (Kosko, 1992). One of the most widely usedneural network topologies is the multilayer perceptron(MLP) because of its applicability to different problems(Hagan, 1997; Schalkoff, 1997). In MLPs, the minimumnumber of layers is three, which includes input, hidden,and output layers. Generally, as the problem getscomplicated, larger and complex ANN topologies maybe required. However, although topologies of this naturegenerally will achieve mostly lower errors, this may alsobe due to over-fitting rather than good modeling(Statsoft, 2003).

The total (net) input to a neuron and its output arecalculated using a transfer function, or axon. This is

suitable network for the development mining simulations without

X-validation Testing

SE Final MSE Minimum MSE Final MSE NMSE R

0.00462 0.00498 0.00498 0.1892 0.903180.00499 0.00621 0.00621 0.2140 0.891370.00312 0.00286 0.00286 0.1142 0.942370.00193 0.00176 0.00176 0.0599 0.969920.00064 0.00106 0.00106 0.0184 0.990850.00184 0.00127 0.00127 0.0656 0.967190.00059 0.00049 0.00049 0.0206 0.989640.00029 0.00023 0.00023 0.0134 0.993350.00062 0.00056 0.00056 0.0219 0.988990.00046 0.00043 0.00043 0.0207 0.989700.00035 0.00034 0.00034 0.0090 0.855490.01201 0.01931 0.01931 0.8952 0.682250.01237 0.02040 0.02040 0.9292 0.775370.01003 0.01681 0.01681 0.7764 0.802880.00574 0.00961 0.00961 0.4895 0.851070.00407 0.00666 0.00066 0.3812 0.844140.00558 0.00937 0.00937 0.4694 0.838810.00202 0.00263 0.00263 0.2690 0.858000.00171 0.00198 0.00198 0.2414 0.874290.00153 0.00167 0.00167 0.2253 0.883220.01249 0.02061 0.02061 0.9434 0.81673

ng the schemes tested. All the learning rules are momentum.


sometimes called a “squashing function” (Eberhart andDobbins, 1990), since it compresses the output rangebetween either 0 to1 or −1 to 1, depending on the choiceof the transfer function. While there are various transferfunctions, the hyperbolic tangent axon (Tanhaxon) andSigmoid functions are generally the non-linear axonsused.

Based on the training method, neural networks areclassified as either supervised or unsupervised net-works (Mohaghegh, 2000). The supervised trainingalgorithm requires repeated showings (Epoch) of bothinput vectors and the expected outputs of the trainingset to the network to let it learn the relations on afeedback basis (Gorucu et al. 2005). The neural net-work computes its output at each epoch and compares itwith the expected output (target) of each input vector inorder to calculate the error. Minimizing the meansquare error (MSE) is the goal of the training process.The most widely used technique is to propagate theerror back and adjust the initially assigned random

Table 6The performance results of different ANN topologies tested to find the suitablshielding boreholes

Hiddenlayers

Neuronperhiddenlayer

Epoch Transferfunction

Momentum Training

Minimum MS

1 1 4 500 Tanhaxon 0.3 0.004122 1 4 500 Tanhaxon 0.5 0.003513 1 4 500 Tanhaxon 0.7 0.003324 1 6 500 Tanhaxon 0.7 0.002115 1 8 500 Tanhaxon 0.7 0.003046 1 6 1000 Tanhaxon 0.7 0.000737 1 6 2000 Tanhaxon 0.7 0.000528 1 6 1500 Tanhaxon 0.7 0.000519 1 8 2000 Tanhaxon 0.7 0.0028210 1 8 2000 Tanhaxon 0.8 0.0003411 1 10 2000 Tanhaxon 0.8 0.0003012 1 10 2000 Tanhaxon 0.7 0.0003413 1 10 1500 Tanhaxon 0.7 0.0007514 1 8 4000 Tanhaxon 0.7 0.0003415 2 6 1500 Tanhaxon 0.7 0.0007616 1 4 500 Sigmoid 0.3 0.0152417 1 4 500 Sigmoid 0.5 0.0156718 1 4 500 Sigmoid 0.7 0.0118619 1 6 500 Sigmoid 0.7 0.0089920 1 8 500 Sigmoid 0.7 0.0090421 1 6 1000 Sigmoid 0.7 0.0046722 1 6 2000 Sigmoid 0.7 0.0022123 1 8 2000 Sigmoid 0.7 0.0023024 1 10 1500 Sigmoid 0.7 0.0025825 1 6 1500 Sigmoid 0.7 0.0025526 1 6 4000 Sigmoid 0.7 0.0019227 2 6 2000 Sigmoid 0.7 0.01506

Data rendered in bold are the most suitable topologies in each category amo

weights to each neuron. This process is called back-propagation. This is the technique used in the modelingdescribed in this paper. Training is one of the mostimportant steps in the development phase of the neuralnetwork, since the weights and the network character-istics will be used later in testing data sets and makingsubsequent predictions.

Cross validation and testing are the performancemeasures of the network with new data that giveconfidence that the network is able to recognize thetraining patterns and generalize the relationships, withina certain degree of acceptable error.

5.2. Reservoir simulations as the source of the input–output patterns

The “dynamic” reservoir simulation model thatpredicted the methane inflows generated “synthetic”case studies to develop an artificial neural network-basedprediction and optimization tool. Airflow requirements

e network for the development mining simulations with degasification/

X-validation Testing

E Final MSE Minimum MSE Final MSE NMSE R

0.00412 0.00567 0.00567 0.13408 0.931070.00351 0.00518 0.00518 0.11492 0.941240.00332 0.00490 0.00490 0.10858 0.944750.00211 0.00289 0.00289 0.06592 0.967330.00304 0.00469 0.00469 0.10479 0.946350.00073 0.00113 0.00113 0.03076 0.984990.00052 0.00085 0.00085 0.02452 0.988450.00051 0.00074 0.00074 0.02174 0.989900.00282 0.00457 0.00457 0.09693 0.950860.00034 0.00059 0.00059 0.01170 0.994200.00030 0.00075 0.00075 0.02005 0.990690.00034 0.00067 0.00067 0.01933 0.990950.00075 0.00082 0.00082 0.01880 0.991080.00034 0.00073 0.00073 0.02162 0.989590.00076 0.00115 0.00115 0.03222 0.985230.01524 0.01171 0.01171 0.95764 0.611470.01567 0.01213 0.01213 0.94738 0.807860.01186 0.00900 0.00900 0.73909 0.852210.00899 0.00657 0.00657 0.55004 0.867020.00903 0.00685 0.00685 0.57277 0.855240.00467 0.00356 0.00356 0.29142 0.891420.00221 0.00184 0.00184 0.14768 0.924870.00230 0.00188 0.00188 0.15540 0.922450.00258 0.00211 0.00211 0.17099 0.911940.00255 0.00211 0.00211 0.17074 0.912330.00192 0.00159 0.00159 0.12652 0.935380.01506 0.01157 0.01157 0.91014 0.83561

ng the schemes tested. All the learning rules are momentum.

Fig. 13. Comparison of target airflow rates (simulator-based predic-tions for the case without degasification wells) of the testing data setwith the predicted rates from the proposed ANN shown in Fig. 12A.

Fig. 12. The ANN topologies developed for two different classes of reservoir simulation for the mine ventilation problems studied in the paper. (A)Prediction of air requirements for ventilating development entries in the absence of any methane degasification and shielding around the entries. (B)Prediction of air requirements in the presence of degasification and the shielding against methane inflow).


were calculated for required methane concentrationsbased on methane inflows predicted by the reservoirsimulation.

For the purposes of developing an ANN tool andtraining it as a predictive tool, the reservoir simulationresults were divided into two groups based on the use ofdegasification. The reservoir simulations without dega-sification resulted in four variables, each with its ownrange of values, and reservoir simulations with thedegasification option resulted in six variables, shown inTable 2. The values of the input variables were changedbetween model runs to create physically meaningfulsimulation data points. Therefore, 360 modeling datapoints (exemplars) for the no-degasification case and720 exemplars for the degasification case weregenerated for these two classes of simulations.

5.3. Development of ANN models

The development and implementation of the ANNwere performed using NeuroSolutions™ version 5.0software (NeuroDimensions, 2006). Entry develop-ments with and without degasification were consideredseparately with two different ANNs constructed for eachcase. The whole data set (exemplars) created for eachclass of problem was separated into three sections to use

them for training, cross validation, and testing purposes.In the case of simulations without shielding boreholes,252 out of 360 exemplars (70%) were saved as trainingdata, 36 exemplars (10%) were saved as cross-validationdata set, and 72 exemplars (20%) were saved for testingthe trained network. In the case of the simulations withshielding wells, the number of exemplars separated fortraining were 288 out of 720 (40%), 144 (20%) for crossvalidation, and 288 (40%) for testing. The criterion inbreaking up the data was to allocate enough exemplarsfor the ANN training and testing.

Fig. 14. Distribution of the number of exemplars within certain relativeerror bins in testing of the completely new data (testing set) for the casewithout degasification wells (Fig. 13) using the network in Fig. 12A.

Fig. 16. Distribution of the number of exemplars within certain relativeerror bins in testing of the completely new data (testing set) for the casewithout degasification wells (Fig. 15) using the network in Fig. 12B.


5.3.1. Performance study for the selection of a suitableANN topology

In order to determine an appropriate ANN topologyand to identify network parameters for training andtesting, various combinations of parameters were tested.Values of minimum and final mean squared error (MSE)were noted for training and cross validation of the ANN.Also, values of nominal mean squared error (NMSE),regression coefficient, and absolute error were noted fortesting the ANN. The number of hidden layers, thenumber of neurons in hidden layers, the number ofepochs, the type of transfer function, and the value ofmomentum were varied. Tables 5 and 6 give the resultsof this parametric search to build an ANN for theanalyzed ventilation problems.

Table 5 gives the results of the parametric search for asuitable network structure function for the ventilation ofentries in the absence of degasification. In thissensitivity study, a single-hidden layer was predictedto be adequate, although one case with two hiddenlayers was tested as well. The best combination of

Fig. 15. Comparison of target airflow rates (simulator-based predic-tions for the case without degasification wells) of the testing data setwith the predicted rates from the proposed ANN shown in Fig. 12B.

network topology and parameter values were obtainedwith 6 neurons, 2000 epochs, and 0.7 as the momentumvalue using a tanhaxon transfer function. The MSEsobtained in the training and cross-validation phases(0.00029 and 0.00023, respectively), and the NMSEand R (0.0134 and 0.993, respectively) were the bestcombinations among the examined networks. Theminimum and maximum absolute errors were 10.3 ft3/min (cfm) (0.00486 m3/s) and 17,400 cfm (8.21 m3/s),respectively. Sigmoid function was also tested withseveral combinations of parameters, but its results werenot as good as the ones obtained with tanhaxon transferfunction (Table 5).

A similar sensitivity analysis was performed with thedata set using the shielding option. Table 6 gives theresults of performances achieved with various combina-tions of network topology and tested parameters. In thisanalysis, 1 hidden layer, 8 hidden-layer neurons, 2000epochs, and a momentum value of 0.8 gave the bestresults for the tanhaxon transfer function tests. Thiscombination gave low MSEs in training and in cross-

Fig. 17. Comparison of monthly-averaged airflow rate measuredduring development of tailgate and headgate entries shown in Fig. 1and the ANN predictions based on the calculated input parameters.


validation runs (0.00034 and 0.00059, respectively), alow nominal mean squared error (0.0117), and a highregression coefficient in the testing phase (0.9942). Theminimum and maximum errors were 4.39 cfm(0.00207 m3/s) and 10,700 cfm (5.29 m3/s), respectively.Although higher numbers of hidden-layer neurons (10)were tested with both 0.7 and 0.8 momentum coeffi-cients, the results did not improve. The Sigmoid functionperformed less efficiently than the tanhaxon function forevery combination tested (Table 6). Thus, the tanhaxonwas selected as the transfer function for the ANN networkto analyze ventilation with degasification.

The topologies shown in Fig. 12 are proposed as atool to analyze and predict ventilation airflow require-ments with and without the use of degasification.

5.3.2. Training of the ANN modelsVentilation of underground coal mines is a complex

process and involves many considerations since themethane gas levels may change. Some of these factorsinclude reservoir properties of the coalbed, miningoperation-related changes in the methane inflow intothe entries, the length of development entries, and thepresence or absence of any previous degasification orshielding. In this study, changes in the reservoir proper-ties of the coalbed were excluded from the variables, byadapting the models for the Pittsburgh Coalbed. Eventhen, it is not easy to predict an optimized ventilationairflow rate to achieve a certain methane level for thethree-entry mine geometry shown in Figs. 1–3. In theabsence of a data library and an appropriate tool toanalyze it, the only choice is to create and run a newreservoir model, which usually is a lengthy process.While it is also possible to make some estimates basedon interpolations, that approach is not only time con-suming but is usually inaccurate. Ultimately, the ANN isexpected to make accurate estimates for completely newinput parameters, by reproducing the results of thesesimulations and approximating the results of a com-pletely new data set.

For training and cross validation of the ANN shownin Fig. 12A, developed for the case without degasifica-tion, 252 training patterns and 36 cross-validationpatterns were shown to the network 2000 times(epoch) to establish the relationships and minimize theerror. The MSE for training and cross validation of thisnetwork as a function of the epoch showed that the errorbegan to decrease rapidly for both training and crossvalidation. However, the minimumMSEs were obtainedat the end of the 2000 epoch cycles. The final MSEsobtained were 0.00029 and 0.00023 for training andcross validation, respectively.

For training and cross validation of the ANNdeveloped for the case with degasification (Fig. 12B),288 training patterns and 144 cross-validation patternswere shown to the network 2000 times as determinedfrom the ANN performance study (Table 6). Thetanhaxon was selected as the transfer function and 0.8was used as the momentum factor for this networkhaving one hidden layer, eight hidden neurons, and sixinput neurons. For this network, the minimum and alsofinal MSEs obtained at the end of the 2000th epochcycle were 0.00034 for training and 0.00059 for crossvalidation.

5.3.3. Testing of the developed ANN modelsThe trained networks were tested in two stages. In the

first stage, the networks were tested using the trainingdata. The same training data set was presented to thenetwork in the testing phase to note the predictionperformance. In the second stage of testing, the ANNwas exposed to a new data set (testing set) that it had notyet seen. This is an important test to see if the ANNgenerates acceptable predictions for the new datapatterns or, in other words, to see whether it recognizesthe patterns and generalizes the results with which it wastrained. In some cases, a trained network may give closepredictions to the training set when the same data is usedin the testing phase. Poor predictive performancetowards a new set suggests that the ANN cannotgeneralize or recognize patterns and this may indicatememorization. Therefore, two-stage testing should givean idea about the performance and thus predictivecapability of the trained ANNs.

In testing of the ANN model without shieldingboreholes (Fig. 12A), the performance of the networkwas first tested against the training data set having 252exemplars. The analysis showed that the airflow ratespredicted by the network were very close to the targetvalues (simulator-based predictions) in the training setfor the majority of the data set. The minimum andmaximum absolute errors were 1.20 cfm (0.00056 m3/s)and 10250 cfm (4.84 m3/s), respectively. The NMSEand R values were 0.0043 and 0.9978, respectively,which shows a good degree of success in the predictionperformance. Using training data set in testing, theprediction errors revealed that the error generallyfluctuated between +/−5% depending on the airflowrate variations between predictions and targets, althoughit is generally less than 5% for most of the data set.Distribution of exemplars within certain error binsrevealed that the error for 50 of the exemplars (19.7%)was between +/−1%. The number of exemplars having+/−3% error was 129 (51%) in the data set. When an


error margin of +/−5% was considered, 183 (72%)exemplars fell within this range. Overall, 91% of thedata had an error less than 10%. The average of theabsolute value of relative errors for the entire set ofexemplars was 4.0%.

In the second stage of testing for this ANN (Fig. 12A),the network was exposed to a new data set (testing set) toevaluate its predictive performance. For this simulation,72 exemplars were saved. Fig. 13 shows the comparisonof airflow rates predicted by the network to the targetvalues of the testing set calculated by the reservoirsimulator. This figure shows that the predictions of thetrained network are close to the simulator predictions forthe completely new data set. For testing with this newdata set, the NMSE and R values were 0.0134 and 0.993,respectively. The minimum and maximum absoluteerrors in airflow rate predictions were 10.3 cfm(0.00485 m3/s) and 17,400 cfm (8.21 m3/s). The numberof exemplars within different relative error bins showsthat the number of exemplars with +/−3% error is 30, or42% of all data (Fig. 14). For 85% of the exemplars, theerror is less than +/−8%. The average absolute value ofrelative error of prediction for the whole set of exemplarsis 4.39%. These results suggest that the network inFig. 13 is capable of making reasonably accuratepredictions for a new data set. This network also looksvery promising for predicting ventilation airflow rates fordevelopment mining.

In testing of the ANN in the presence of shieldingboreholes (Fig. 12B), the same procedure was applied.ANN learning was tested against the training data set.For this process, a training data set with 288 exemplarswas introduced to the network as the testing, or “new,”data set. This analysis showed that the predicted valueswere close to the target values of the data set. Thenominal MSE and R values were 0.0076 and 0.9962,respectively. The low error and high correlationcoefficient are good indicators of predictive perfor-mance. The minimum and maximum absolute errors inairflow rates for the predictions were 11.4 cfm(0.00537 m3/s) and 10920 cfm (5.15 m3/s), respectively.In this phase, the relative errors using the predicted andtarget airflow rates showed that the error fluctuatedwithin a narrow band. For most of the exemplars, theerror was +/−10%. The average of the absolute values ofthe relative errors was 7.0%. The distribution ofexemplars within different error bins shows that 137(48%) of all exemplars had a relative error +/−5%. Thenumber of exemplars with less than +/−10% relativeerror was 222 (77%). For a +/−15% error range, thenumber of exemplars within this range of error was 258,or 89% of all data.

In the second stage, the performance of this ANN(Fig. 12B) was assessed with a new data set containing288 exemplars (testing set). The result indicatedreasonably good predictive capability for the new dataset and show that the network can recognize andgeneralize patterns of input data. The minimum andmaximum absolute errors between predictions and targetsare 4.4 cfm (0.002 m3/s) and 10,660 cfm (5.03 m3/s),respectively with an NMSE of 0.00117 and an R value of0.9942 (Fig. 15). The error histogram in Fig. 16 showsthat 85 of the exemplars (30% of the entire set) and 133 ofthe exemplars (46% of the set) have less than +/−3 and+/−5% error, respectively. This data also shows 208exemplars (72%) with +/−10% error and 27 exemplars(9%) with an error of +/−15% or more.

The comparison of ANN predictions for the airflowrates and the error analyses show that the proposedANNs perform reasonably well for the prediction andoptimization of ventilation air requirements. The flexi-bility and predictive capability of trained ANNs offeradvantages in decision making and design processes.

5.4. Comparison of airflow rate predictions of thetrained ANN with the in-mine measurements

The trained ANN without shielding boreholespredicted ventilation airflow rates during developmentof tailgate and headgate entries shown in Fig. 1. Fig. 17shows the measured airflow rates, reported as monthlyaverages, and the predicted airflow rates using thetrained ANN shown in Fig. 12B. The following inputdata (Tables 3 and 4) was collected and used by theANN to predict airflow rates: monthly average methaneconcentration data measured in the mine, calculatedmining advance rates using reported monthly linearmining distances, calculated average mining heightusing reported monthly coal production tonnages andthe dimensions of the entries, and the calculated entrylength based on the monthly linear footage mined. Asshown in Fig. 17, the predictions of the trained ANN aregenerally in good agreement with the field data. Thereare a few data points that show relatively high errorcompared to others, which may have arisen from thecalculated parameters that may have been different fromtheir actual field values.

Unfortunately, the predictions of the ANN withshielding boreholes could not be compared with in-minemeasurements because of the absence of reported data.However, it would be reasonable to expect acceptableagreement with predictions and field data. One of theadvantages of developing ANNs in conjunction withrealistic simulations is that it increases the predictive


versatility of the ANN when other data sources arelimited.

Although the reservoir and the ANN models fordevelopment mining were based on Pittsburgh Coalbedparameters, the values of the coalbed reservoir proper-ties may be varied in the parametric reservoir simula-tions. Developing and training potentially larger ANNmodels would enable the mining and the coalbedparameters to be generalized for operations in othercoal regions.

6. Summary and conclusions

This study presented the development and applica-tion of “dynamic” reservoir models and artificial neuralnetworks to predict and optimize methane inflows andventilation airflow requirements for development min-ing of coal seams. The reservoir models were developedfor a typical three-entry system practiced in the NorthernAppalachian Basin section of the Pittsburgh Coalbed.These models considered the presence and absence ofboreholes against methane inflow. Model simulationswere performed by changing various mining anddegasification parameters in a systematic manner topredict methane inflows and airflow requirements. Thereservoir model predictions were compared with the in-mine measurements available from the mining oftailgate and headgate entries around a longwall panelin the Pittsburgh Coalbed.

The predictions from the reservoir simulations wereused to develop, train, and test the artificial neuralnetworks (ANNs). Various network structures weretested to evaluate the predictive capabilities of the ANN.The ANN predictions were compared with the reservoirsimulation results and the resulting prediction errorswere evaluated. Finally, the ANN was asked to predictthe airflow rates measured in the entries of thePittsburgh Coalbed mine using reported and calculatedmining parameters.

This study made the following conclusions:

1. Reservoir simulation can be used effectively formodeling development mining in underground coalmines. It can predict methane inflows into the entriesbased on various coalbed and operating parameters,from which the ventilation airflow requirements canbe predicted to maintain a desired methane level.

2. Methane inflow rate increases with entry lengthbecause of increases in the surface area of thecoalbed exposed to the mine environment. Theseincreases are a logarithmic function of the miningrate. However, the effect of the mining rate is less

pronounced on methane inflow for shorter devel-opment distances compared to longer developmentdistances.

3. For shorter development distances, marginal adjust-ments in ventilation airflow may be adequate to keepmethane levels under 1%. The model showed thatthis adjustment was generally independent of themining rate. However, longer developments requiresignificantly more ventilation air capacity to keepmethane levels constant. This amount is a function ofthe mining rate.

4. Employing shielding boreholes to protect entriesfrom methane migration during mining is an effectiveapproach. Even if the boreholes cannot be operatedfor a long time prior to the start of mining, theirpresence during mining makes a big difference (about25%) in methane inflow rates, especially duringmining of longer sections.

5. Positioning of boreholes relative to entries is impor-tant. These results show that positioning wellbores asclose to the entries as practically possible is moreeffective, especially when mining long developmentsat higher mining rates.

6. Parametric and realistic reservoir simulations candevelop, train, and test ANN-based prediction andoptimization models, where actual field data isunavailable. The ANNs developed in this study cangenerate the reservoir simulation data with areasonable accuracy for predicting ventilation airrequirements during development mining.

7. ANN predictions of in-mine measured airflow datawere successful when using calculated monthly-averaged input parameters for predictions rather thanthe spatio-temporal data.

8. The ANN can be generalized to other coalbeds byusing parametric reservoir simulations to define theimpacts of different coalbed parameters on airflowrequirements and methane inflows.

References

Brunner, D.J., Schwoebel, J.J., Li, J., 1997. Simulation baseddegasification system design for the Shihao Mine of the Songzaocoal mining administration in Sichuan China. Proc. 6th Interna-tional Mine Ventilation Congress, Pittsburgh, PA, pp. 429–433.

Computer Modeling Group Ltd., 2003. Generalized Equation of StateModel-GEM. Calgary, Alberta, Canada.

Diamond, W.P., 1994. Methane control for underground coal mines.Information Circular No. 9395, US Dept. of Interior, US Bureau ofMines.

Diamond, W.P., La Scola, J.C., Hyman, D.M., 1986. Results of direct-method determination of the gas content of the U.S. coalbeds.Information Circular No. 9067, US Dept. of Interior, US Bureau ofMines.


Eberhart, R.C., Dobbins, R.W., 1990. Neural Network PC Tools: APractical Guide. Academic Press Inc., San Diego, CA.

Ertekin, T., Sung, W., Schwerer, F.C., 1988. Production performanceanalysis of horizontal drainage wells for the degasification of coalseams. Journal of Petroleum Technology 625–631.

Gorucu, F.B., Ertekin, T., Bromhal, G., Smith, D., Sams, N., Jikich, S.,2005. Development of a neuro-simulation tool for coalbedmethane recovery and CO2 sequestration. 2005 InternationalCoalbed Methane Symposium, Tuscaloosa, AL. Paper 0505.

Hagan, M.T., Demuth, H.B., Beale, M.H., 1997. Neural NetworkDesign. PWS Publishing, Boston, MA.

Hardcastle, S.G., Dayss, A., Leung, E., 1997. Integrated mineventilation management system. Proceedings of 6th InternationalMine Ventilation Congress. Pittsburgh, PA.

Karacan, C.Ö., Diamond, W.P., Esterhuizen, G.S., Schatzel, S.J., 2005.Numerical analysis of the impact of longwall panel width onmethane emissions and performance of gob gas ventholes. 2005International Coalbed Methane Symposium. Tuscaloosa, AL.Paper 0505.

Karacan, C.Ö., Esterhuizen, G.S., Schatzel, S.J., Diamond, W.P., (inpress). Reservoir simulation-based modeling for characterizinglongwall methane emissions and gob gas venthole production.International Journal of Coal Geology.

King, G., Ertekin, T., 1991. State of the art modeling for unconventionalgas recovery. SPE Formation Evaluation 63–72 (March).

Kosko, B., 1992. Neural Networks and Fuzzy Systems: A DynamicalSystems Approach to Machine Intelligence. Prentice-Hall, Inc.,Englewood Cliffs, NJ.

Mohaghegh, S., 2000. Virtual-intelligence applications in petroleumengineering: Part 1-Artificial neural networks. Journal of Petro-leum Technology 52 (9), 64–73.

NeuroDimension, 2006. NeuroSolutions 5.0.Gainesville. FL.Noack, K., 1998. Control of gas emissions in underground coal mines.

International Journal of Coal Geology 35, 57–82.Schalkoff, R.J., 1997. Artificial Neural Networks. McGraw-Hill, New

York.Statsoft., 2003. Neural Networks. http://www.statsoftinc.com/textbook/

stneunet.html.Thakur, P.C., Davis, J.G., 1977. How to plan for methane control in

underground coal mines. Mining Engineering 41–45 (October).Thakur, P.C., Poundstone, W.N., 1980. Horizontal drilling technology

for advance degasification. Mining Engineering 676–680 (June).Thakur, P.C., 1997. Methane drainage from gassy mines—a global

review. Proc. 6th International Mine Ventilation Congress.Pittsburgh, PA, pp. 415–422.

Young, G.B.C., Paul, G.W., Saulsberry, J.L., Schraufnagel, R.A.,1993. A simulation-based analysis of multiseam coalbed wellcompletions. 68th Society of Petroleum Engineers AnnualTechnical Conference and Exhibition. Paper No 26628.

Zuber, M.D., 1997. Application of coalbed methane reservoirsimulators for estimation of methane emissions in longwallmining. 6th International Mine Ventilation Congress. Pittsburgh,PA, pp. 436–440.

Zuber, M.D., 1998. Production characteristics and reservoir analysis ofcoalbed methane reservoirs. International Journal of Coal Geology38, 27–45.

development and application of reservoir models and

Documents